Computing and maths has the smallest protected share of any group in the economy. Seven of its 36 occupations have no protected work at all. This is the chapter everybody else should read.
Exposure means published research judged an AI could halve the time a task takes, at the same quality. It is not a prediction about any job. How to read this
Across this group, 9 per cent of tasks cannot be sped up by an AI. That is the lowest figure in the economy. In work needing the least preparation the same figure is 76 per cent, and in construction it is 90.
The people who build these systems have less shelter from them than anybody. That is not irony, it is the gradient in the overview working exactly as described: this is text-shaped, screen-based, highly trained work, which is the profile the ratings reward.
Not a small amount. None. Every task listed for Data Warehousing Specialists, Document Management Specialists, Information Security Analysts, Mathematicians, Operations Research Analysts, Video Game Designers and Web and Digital Interface Designers was rated as something an AI could halve, either today or with a tool built on top.
Those lists run from 11 to 30 tasks, so a zero means no task on a short list was rated protected rather than that none could exist. The direction is clear even where the precision is not.
Sixty-two per cent of the work here sits in the band that needs purpose-built software. Everywhere else in this report that band is hypothetical: somebody judged a tool could be built and nobody has built it.
Here they have. Coding assistants are a shipped product category with millions of users, aimed squarely at the tasks in that band. Which makes this group a preview rather than a warning.
Software Developers is the largest occupation here at 1.36 million people, and reads as barely touched: 5 per cent reachable with a chat window, 79 per cent needing a tool. Anybody who has watched software development since 2023 knows that is no longer true, which is the clearest illustration in this report of why every figure in it is a floor.
Solid is reachable with a chat window today, pale needs software building first, grey cannot be sped up. Task counts are shown because several of these rest on fewer than twenty.
Every other group has a large middle band and almost no tools built against it. The People profession's is 55 per cent. Computing and maths shows what that band looks like once somebody builds: not the job disappearing, but the job changing shape faster than a 2023 rating can describe.
That is the argument for reading the middle band as a backlog rather than a forecast. Somebody built the tools for this group because the people in it could build them.
Not because they are cleverer about AI, but because somebody built the tools for their middle band and nobody has built yours. The gap is not capability. It is who happened to be able to build.
Worth knowing before the next conversation about whether People teams are behind on AI. On this evidence they are not behind. They are under-tooled, which is a different problem with a different fix.
There is no data here on what anybody actually does with AI. A dataset that appears to measure it was removed from this report because its meaning could not be stated. The reasoning is on the limits page.
The exposure ratings cover all 771 tasks in these 36 occupations with no gaps, and were made in 2023 against the model capabilities of that year. Read every figure as a floor.
Exposure ratings from Eloundou, Manning, Mishkin and Rock, GPTs are GPTs: Labor market impact potential of LLMs, Science 384, 1306–1308, 2024, used under the MIT licence. This page includes information from the O*NET Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), used under the CC BY 4.0 license. O*NET® is a trademark of USDOL/ETA. People Team AI has modified all or some of this information. USDOL/ETA has not approved, endorsed, or tested these modifications. Employment figures from the US Bureau of Labor Statistics.